2.3.1. Parameters

The FLUS model is employed to realize the spatial layout of PLE land use. The detailed theory and process can be found in Liu's article [41]. Parameters mainly used in the model are described as follows:

#### (1) Suitability probability

The suitability probability of different land types on each parcel can be calculated based on the powerful, intelligent prediction function of the FLUS model's neural network by inputting the historical data of driving factors related to land-use evolution. The amount of future land demand is determined by other methods. In this study, SD is coupled with FLUS to simulate the spatial layout of the PLE space of Zhaotong city in 2030.

The selection of driving factors is crucial to the ANN-based suitability probability. This paper references previous literature studies [52–54]. Sixteen driving factors of different aspects of land use were taken into account, such as topography (including elevation, slope, and slope direction), natural meteorology (including soil conditions, precipitation, and annual mean air temperature), and social economy (including population, GDP, distance from the city center, commuting time). Data sources are shown in Table 3. All raster datasets were resampled to 30 m by the software Arcgis 10.2.

**Figure 4.** Interrelationship diagram of the PLE system dynamics model. Variables in underlined italics are control variables.



(2) Conversion cost matrix

The cost matrix reflects the conversion rules between different land types, in which 0 means that one type of land is not allowed to be converted to another, 1 means that it can be converted. The conversion cost matrix for different land types in this paper varies during model validation and scenario simulation.

(3) Neighborhood factor values

Referring to the method of Ou et al. [55] and expert empirical knowledge, the neighborhood weights of the nine land types in this paper were determined, as shown in Table 4. The weight values range from 0 to 1. A value closer to 1 represents a stronger expansion capacity.

**Table 4.** Neighborhood factor values for different PLE subclasses.


#### (4) Other parameters

Another parameter is the time t in the formula, representing the number of iterations. More iterations mean more time and lower efficiency. To fully carry out spatial simulation and take simulation efficiency into account at the same time, the number of iterations is set as 600 and the acceleration factor as 0.5.

#### 2.3.2. Verification

The feasibility and effectiveness of the FLUS model should be verified at first. In this paper, land use data from 2010 and 2015 are used as the base period data to simulate those of 2015 and 2018, respectively.

Firstly, the suitability probabilities of different land types in 2010 and 2015 were calculated using artificial neural networks (ANNs). Accuracy can be measured by the indicators RMSE (root mean square error), ROC (receiver operating characteristic), and AUC (area under curve) [41]. Smaller RMSE values and larger AUC values indicate higher model accuracy. Results showed that the RMSE was 0.2464 in 2010 and 0.2471 in 2015. AUC values corresponding to the ROC curves are shown in Table 5. ROC curves are shown in Figures S1 and S2.

**Table 5.** AUC values based on ANN suitability probability calculation results.


Secondly, conversion cost matrices used for verification are obtained from the transition matrix between the years 2010~2015 and 2015~2018. If there is a conversion happened between two land classes during the above period, it is assigned as 1; otherwise, it is assigned as 0. They are shown in Tables S3 and S4, respectively. Other parameters are the same as those described in the Parameters section.

Results showed that compared with the actual data, the kappa coefficient (an indicator used to measure accuracy) of the 2015 simulation is 0.91, with an overall accuracy of 93.35%. The confusion matrix (a specific table layout that allows visualization of the performance of an algorithm in the field of machine learning) between the actual data and simulation is shown in Tables S5 and S6. Similarly, the kappa coefficient of the 2018 simulation is 0.85, with an overall accuracy of 88.99%. Both kappa coefficients are greater than or equal to 0.85, indicating that the FLUS model has high simulation accuracy and can be used for future simulations.
